Broadcom to lend Anthropic up to $42bn for TPUs as US datacenter buildout hits packaging ceiling

The loan could cover about a third of the $125.2bn Anthropic has committed to a five-year lease of tensor processing units, the AI chips Google designs with Broadcom, according to Reuters, which reviewed the filing. The money would come through convertible notes that could convert into Anthropic shares. Anthropic said in the filing it does not expect any notes to be sold before its IPO. Broadcom can also name a financing partner, the report said.

Anthropic declined to comment and Broadcom did not comment, Reuters said. TNW has not seen the prospectus.

The filing itself flags the arrangement's awkward geometry. Anthropic says Broadcom's double role, as hardware supplier and financing partner, creates "potential conflicts of interest" that could affect its access to computing power. It also warns that some payment or performance defaults could make much of its lease obligations due at once. In that case it could draw only limited amounts from the $42bn facility to cover them. Anthropic put cash into a restricted account for Broadcom's benefit in April 2026, according to the filing, and may have to add more in certain circumstances.

Reuters frames the deal as a mirror of Nvidia, which has used its balance sheet to support chip sales. The Financial Times reported this week that Nvidia had talked to insurers about loans backed by its chips. "Nvidia is putting in place a massive amount of its balance sheet, and Broadcom is having to follow suit," Seaport Research analyst Jay Goldberg told Reuters. Robert Leitao, managing partner of Rothschild & Co, told the agency: "It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened."

Anthropic lost $42bn in 2025 as revenue grew 12-fold, its prospectus shows, and its IPO could value it at $2tn, Reuters reported. Broadcom projects AI chip revenue of about $115bn in fiscal 2027 and $230bn in fiscal 2028, and Reuters says Anthropic is on course to become Broadcom's largest compute customer that year.

Packaging becomes the bottleneck

The money lands against a physical constraint. A Jefferies report shared with The Register on 30 September cites analytics firm SynMax, which uses weekly satellite imagery to track land clearing and construction at datacenter sites, and estimates US deployments will rise from roughly 11 GW of gross capacity in 2025 to 16-18 GW in 2026, above its previous forecast of 14-16 GW. It puts the practical upper limit for 2027 in the low 20s of gigawatts.

Land clearing for new projects has plateaued, The Register reported, leaving visible activity well short of the more than 80 GW implied by some announced pipelines and chip demand models for 2028. A previous Jefferies report found only half the US capacity scheduled for 2026 was under construction, with work yet to begin on as much as 80 percent of the 2028 pipeline.

Advanced packaging, where accelerator dies are combined with high-bandwidth memory before OEMs can install them, is the second brake. SynMax estimates existing packaging capacity supports accelerators drawing roughly 13 GW of gross power, equal to about 17.5 GW of total datacenter power after CPUs, memory and cooling. Two packaging projects due online in 2027 could add 6 GW, keeping the ceiling in the low 20s unless capacity expands faster than forecast.

Some of that capacity is shifting. Shinko Electric has acquired Japan Display's Mobara plant and is converting it into an AI semiconductor packaging base, according to finance.biggo.com, a move Tech Times covered on 30 September as a glass substrate push.

China builds a parallel stack

Huawei's rotating chairman Eric Xu claimed this week that the company's Ascend neural processing units have surpassed Nvidia in Chinese market share. "It's pretty hard to collect data about the market share of Nvidia in China, but based on the data we have collected, Ascend has surpassed Nvidia," he said in a Q&A release covered by The Register on 30 September. Nvidia executives said on their Q2 earnings call that H200 shipments to China amounted to less than 1 percent of datacenter revenues.

DeepSeek open-sourced six software modules for Ascend on Wednesday, including an Ascend-compatible version of TileLang, a programming language developed at Peking University that it says is simpler than Nvidia's CUDA, according to the South China Morning Post and The Decoder. Huawei said it "fully supported" the work, and the two companies optimized a supernode of 128 Ascend 950 chips.

Neither side is standing still on design tooling. OpenAI and Synopsys signed a multi-year partnership to build GPT-Synopsys, a model intended to reason about chip design and operate Synopsys' EDA tools, with early tests at semiconductor customers already underway, The Decoder reported on 30 September. Tom's Hardware, running an AI chip design week from 28 September to 2 October, notes that Cadence, Synopsys and Siemens all claim varying levels of agent autonomy, but that every speed figure so far is the vendor's or a customer's own, with "up to" and "early evaluation" caveats and different baselines.